Finger Vein Recognition Based on Multi-Orientation Weighted Symmetric Local Graph Structure

Abstract

Finger vein recognition is a biometric technology using finger veins to authenticate a person, and due to its high degree of uniqueness, liveness, and safety, it is widely used. The traditional Symmetric Local Graph Structure (SLGS) method only considers the relationship between the image pixels as a dominating set, and uses the relevant theories to tap image features. In order to better extract finger vein features, taking into account location information and direction information between the pixels of the image, this paper presents a novel finger vein feature extraction method, Multi-Orientation Weighted Symmetric Local Graph Structure (MOW-SLGS), which assigns weight to each edge according to the positional relationship between the edge and the target pixel. In addition, we use the Extreme Learning Machine (ELM) classifier to train and classify the vein feature extracted by the MOW-SLGS method. Experiments show that the proposed method has better performance than traditional methods.

1. Introduction

Biometrics [1] [2], are identity authentication methods using physical characteristics, such as those of the face, fingerprints, irises, and veins, and so on. Compared with traditional features, such characteristics are safe, unique, lifelong invariant, and therefore widely used. Biometric systems typically include registration and authentication two stages.

In 1992, researchers in Hokkaido University first demonstrated that no vein characteristics are exactly the same for any two people in the world, even identical twins [3]. Finger vein authentication [4] [5] offers more security, stability, universality, liveness, and non-contact than a fingerprint test. It also has advantages over hand vein and palm vein tests: it requires a smaller volume; it has higher credibility and adequate sources of information [6]; and it offers useful redundancy: after the information entry registration for multiple fingers, if one finger is injured, we can still test the others. However, finger vein recognition is easily influenced by light condition [7], scattering [8], ambient temperature [9], and other factors.

For finger vein recognition, feature extraction is one of the key steps. Subspace methods are widely used. In 1991, Turk et al. [10] proposed using eigenfaces (also called Principal Component Analysis, PCA) for face recognition, and it is used for finger vein recognition [11] too. Linear Discriminant Analysis (LDA) [12], Two Dimensional Principal Component Analysis (2DPCA) [13], and Independent Component Analysis (ICA) [14] algorithms have also been proposed. In 1996, Ojala et al. [15], proposed a LBP algorithm and introduced the local information to feature extraction, which measured and extracted the local texture information of image with illumination invariance, and greatly improved the efficiency of facial recognition. However, the LBP operator was unable to extract large texture features. Ojala et al. [16] thus improved it by extending the 3×3 neighborhood to any neighborhood, and using the round neighborhood instead of the square neighborhood. The circular LBP operator only considers 8 neighborhood pixels, which makes use of the surrounding pixels deficient; therefore, in 2011, Rosdi et al. [17] extended the LBP operator in two directions, making it an LLBP operator, with more efficient feature extraction. They applied LLBP to finger vein recognition. However, LLBP only extracts horizontal and vertical line patterns. Therefore, Lu et al. [18] proposed the Generalized Local Line Binary Pattern (GLLBP) method, using a circular neighborhood of multi-direction feature extraction, and applied it to the vein recognition. In 2012, Meng et al. [19] proposed a Local Directional Code (LDC) feature extraction method to finger vein recognition by coding the image with the gradient information. Yang et al. [20] proposed a personalized best bit map (PBBM) algorithm and applied it to finger vein recognition. The PBBM is based on LBP, and in the matching phase, it only uses the best bits to improve recognition performance. In order to solve the common rotation problem of the finger vein images, Pang et al. [21] applied the SIFT feature extraction method, with good results. Xie et al. [22] proposed a Guided Gabor filter to further improve the extraction of finger vein features.

In 2011, inspired by the dominating set, Abusham et al. [23] first proposed the Local Graph Structure (LGS) algorithm and applied it to face recognition, which improved results resisted the influence of light. In 2014, Mohd et al. [24] proposed a symmetrical LGS (SLGS) algorithm. Compared with the LGS algorithm, the SLGS algorithm utilized the spatial information of the pixels in balance. However, SLGS has some deficiencies. When the SLGS operator assigns weight for each edge, it does not take the positional relationship between the edge and the target pixel into account, and the weight of the left pixels is too large, which does not take full advantage of the spatial information of pixels around the target pixel. In addition, the SLGS operator only uses the zero-degree direction information of the target pixels, not the location and gradient information of the surrounding pixels in the other direction.

To overcome the shortcomings of traditional methods, this paper proposes a new method named Multi-Orientation Weighted Symmetric Local Graph Structure (MOW-SLGS), and uses the Extreme Learning Machine (ELM) to train and classify the feature. In the neighborhood of 5 × 5, we not only use the relationship between the target pixel and the surrounding pixels, but also the relationships between surrounding pixels. Moreover, we extract the features in 0 degree, 45°, 90°, and 135° direction by taking advantage of location relationship between pixels and gradient relationship and making full use of surrounding pixels. ELM [25] is a single hidden-layer feedforward neural network, fast and effective at classification. Experiments show the proposed method has better performance than traditional methods.

The second part of this paper introduces the theories of LGS and SLGS; the third part demonstrates the proposed method; the fourth part shows the experimental results, and the fifth part gives the conclusion.

 

2. Related Theory

2.1 LGS

The LGS method can weaken the influence of light on efficiency and improve calculation efficiency, which can be applied to the real-time systems. When assigning weights to the target pixel, it considers all relationships between pixels, not just those which directly include the target pixel, which improves recognition.

The LGS [26] algorithm is dominated by the dominating set in graph structure. Given an undirected graph G = , where V is the point set whose size is n and E is the set of edges, then S, a subset of V, can be called a dominating set if and only if for any point v in V-S, there is a point u of S and (u, v) ∈ E. That is, for any vertex v in G, if v either belongs to D or is adjacent to a vertex in D, then D is called a dominating set of G.

For each target pixel, the algorithm selects 5 surrounding pixels and the composition of the surrounding structure is as shown in Fig. 1. Starting from the target pixel, at the left side of the target pixel, the algorithm compares the values of pixels one by one, counterclockwise, and if the pixel in the direction of the arrow’s value is greater than the previous pixel value, then the connection between the two pixel sides is set to 1. Otherwise, it is set to 0. To the right of the target pixel, values are also compared one by one in a clockwise direction.

Fig. 1.LGS operator graph structure

Through the above process, we can obtain 8 binary values, which are combined into a binary string to obtain the final value of the target pixel. As shown in Fig. 2, we calculate the feature value of the target pixel with LGS operator, like so:

Feature(00110100)=0×128+0×64+1×32+1×16+0×8+1×4+0×2+0×1=52

Fig. 2.Process of LGS operator calculations

2.2 SLGS

The SLGS algorithm [24] for texture-based image recognition has remarkably improved performance. It makes better use of spatial information than other feature extraction methods, and uses a more balanced relationship between surrounding pixels than LGS.

In Fig. 2, the target pixel used the two pixels on the left side and three pixels on the right side of the target pixel. This asymmetry is a problem with LGS. SLGS algorithm, as shown in Fig. 3, adds the idea of a symmetric algorithm into LGS, using symmetrical structure to express the relationship between adjacent pixels, making better use of the spatial information between the image pixels for the target pixel. It symmetrically selects 6 surrounding pixels, starting from the target pixel; at the left side of the target pixel, the algorithm compares the values of pixels one by one counterclockwise as before, and to the right, clockwise as before. As shown in Fig. 4, we calculate the feature value of target pixel with SLGS operator like so:

Feature(00101010)=0×128+0×64+1×32+0×16+1×8+0×4+1×2+0×1=42

Fig. 3.SLGS operator graph structure

Fig. 4.Process of SLGS operator calculations

By comparing the magnitude between pixels adjacent to each other and giving different weights, we obtain a component of 8 binary strings and convert it to an integer, which is the target pixel.

 

3. Proposed Method

3.1 MOW-SLGS

When the SLGS operator assigns weight for each edge, it does not take into account the positional relationship between the edge and the target pixel, denying itself full advantage of the spatial information of pixels around the target pixel. The weight of the left pixels is too large. To balance the weight of right and left sides, this paper assigns weight to each edge according to the position relationship between the edge and the target pixel. We balance the weight of pixels on the right side and left side of the target pixel, and multiply the edge’s weight by distance between surrounding pixel and target pixel. The Weighted SLGS (W-SLGS) operator is shown in Fig. 5.

Fig. 5.The design of weights for Weighted SLGS

In addition, because the SLGS operator only uses the 0° direction information of the target pixels, taking location information and direction information between the image pixels into account, this paper proposes an MOW-SLGS operator.

The proposed method is as shown below.

As shown in Fig. 6 and Fig. 7, we do SLGS operations for the target pixel in the 0°, 45°, 90°, and 135° directions. In the 45° and 135° direction, as shown in Fig. 8, starting from the target pixel and going to the upper left or lower left, the algorithm compares the values of pixels one by one, counterclockwise, as before. The other two angles compare in a clockwise direction as before.

Fig. 6.MOW-SLGS operator graph structure in 45° and 135° direction

Fig. 7.MOW-SLGS operator graph structure in 0° and 90° direction

Fig. 8.Calculation process of MOW-SLGS operator in 45° and 135° direction

In 0°direction, the comparative method is the same as that of SLGS. In 90° direction, as shown in Fig. 9, starting from the target pixel, and going up, the algorithm compares the values of pixels counterclockwise as usual, and going down, clockwise as usual.

Fig. 9.Calculation process of MOW-SLGS operator in 0° and 90° direction

After processing with the MOW-SLGS operation, we can obtain 4 feature values: Feature0, Feature45, Feature90, and Feature135. The bigger the change in the direction feature value, the bigger the change in pixels. Therefore, the direction can better reflect the position information and gradient information of pixels around the target pixel. So in this paper, we choose the maximum value as the feature value of the target pixel. As shown in Fig. 8 and Fig. 9, we calculate the feature value of target pixel with MOW-SLGS operator as follows:

Feature0(01001100)=0×128+1×64+0×32+0×16+1×8+1×4+0×2+0×1=76

Feature45(00110101)=0×128+0×64+1×32+1×16+0×8+1×4+0×2+1×1=53

Feature90(00011100)=0×128+0×64+0×32+1×16+1×8+1×4+0×2+0×1=28

Feature135(01100011)=0×128+1×64+1×32+0×16+0×8+0×4+1×2+1×1=99

Feature=max{Feature0, Feature45, Feature90, Feature135}=Feature135=99

In the MOW-SLGS operator, we set the weight according to the distance between pixels, and the closer the distance, the greater the impact on the target pixel, so the weight we assign corresponds with the image distribution in space. And the MOW-SLGS operator makes full use of position and gradient information, making it more resistant to changes in light, temperature, and other external conditions. Additionally, we extend the operator in four directions, enriching the biometrics and improving feature extraction.

3.2 Dimensionality Reduction

To extract features better and remove redundancy and noise from finger vein features, we use the PCA algorithm [10] to reduce the dimensionality of the feature matrix. When the contribution rate is 0.95, the main feature extracted by PCA algorithm can already describe the overall characteristics, while the characteristic dimension is not large, and the recognition performance is not very different from the performance when contribution rate is greater than 0.95.In the experiment, we select 0.95 as the contribution rate.

3.3 Training and Classification

In our experiments, we use the ELM [26] for training and classification.

A standard single-hidden-layer feedforward neural network which contains N training samples {xi, yi} (i = 1, ..., N), Ñ hidden-layer nodes and has excitation function g(x) can be expressed as follows:

Where, ωi represents the input weights of the input neuron and the i-th hidden-layer node; βi is the output weights of i-th hidden-layer nodes and output neurons; bi is the bias of i-th hidden-layer node; oj represents the output value of j-th sample.

The process of ELM algorithm is as follows:

Step 1: Set the input random weights ωi, and bias bi, i = 1, ...,Ñ ;

Step 2: Calculate the hidden-layer output matrix H;

Step 3: Calculate the output weights β:

Compared with other training algorithm, ELM algorithm does not need to adjust the input weights and the bias of hidden network neurons, and we can get a unique optimal solution at higher efficiency.

In our experiments, we set the number of hidden neurons to 2000, the reason will be given in Section 4.

 

4. Experiment and analysis

4.1 Experimental database

The database used here is the open Homologous Multi-modal Traits Database [28] developed by Shandong University. During finger vein image acquisition, each person must provide the index finger, middle finger, ring finger of both hands. Each finger produces 6 images, meaning 36 images in total for one person. In this database, there are 106 persons, each with 36 images, for a total of 3816 images. The size of each image is 320 × 240 as shown in Fig. 10.

Fig. 10.Image of finger vein database

In our experiments, we select 100 different fingers from the original database, and each finger has 6 images, so we have a total of 600 images in the new database. We first crop the ROI images(Fig. 11), then use the Guided Gabor enhancement method on them [22], which yields image in the size of 256×96 (Fig. 12).

Fig. 11.Finger vein images after ROI extraction

Fig. 12.Finger vein images after preprocessing

All the experiments are implemented with MATLAB 7.0, and performed on a PC with a 2.8 GHz CPU and 2.0 G memory in windows 7 OS.

4.2 The experimental results

4.2.1 Determine the number of hidden neurons

We conducted the experiment to test the influence of hidden neurons on the performance of finger vein recognition system, and the result is shown in Table 1. For each finger, we used four images for training, two for testing. From the following table, we can see when the number of hidden neurons is more than 1500, the recognition performance tends to be stable at about 0.9100. So, in order to save time, we choose 2000 as the number of hidden neurons.

Table 1.Recognition rate of different number of hidden neurons

4.2.2 Determine the best image blocking method

As the size of the feature matrix depends on the number of block images processed, we investigate the effect of image blocking method on the experimental results. We first explore the best image blocking method. The image blocking method is the method to block a finger image to obtain the sub-images. The blocking method M × N means that we divide the image into M rows and N columns.

Since the number of sub-images determines the size of the image feature matrix, but when the number of sub-images is too large, the feature matrix is too big, it will cost a lot of time. Therefore, in the trial, we selected the following blocking methods from the original image: 1×1, 2×1, 2×2, 4×2, and 4×4. The experimental results are shown below. For each person in the above database, we used three images for training, the remaining three for testing, and the MOW-SLGS algorithm for testing.

Fig. 13 and Table 2 show that, among different blocking methods, the recognition rates fluctuate slightly, but 4×2 produces the best rate; therefore, we use the 4×2 blocking method. Rates for 1×1, 2×1, 2×2, 4×2, and 4×4 blocks are 86.67%, 86.50%, 86.33%, 87.17%, and 87.00%, respectively.

Fig. 13.Recognition rate of different block methods

Table 2.Recognition rates of different blocking method

4.2.3 Comparison of performance using different classifiers

In our experiments, we tested the performances by using three state-of-the-art classifiers: the Nearest Neighbor (1-NN), ELM, and VBELM [28]. Again we used four images for training, two for testing, and the MOW-SLGS algorithm for testing. In the training stage, we also use MOW-SLGS for feature extraction. Each experiment was repeated with 20 times and chose the average value as result. The results are shown as Table 3.

Table 3.Recognition rates of different classifiers

From Table 3, we can see that the 1-NN method does not train the samples, so we just compare the training time between ELM and VBELM. When the number of training samples is 4, the training times for ELM and VBELM are 1.8188 s and 111.8531 s and the testing time for 1-NN, ELM, and VBELM are 26.4529 s, 0.0938 s, and 0.1008 s, respectively. While the recognition rate of 1-NN, ELM, and VBELM are 85.00%, 90.85%, and 91.25%, respectively. The training time of VBELM is almost 60 times that of ELM, and the recognition rate is lower than ELM. In addition, the recognition rate of ELM is higher than 1-NN, and its test time is lower than 1-NN’s; therefore, this paper chose ELM as the classifier for training and classification.

4.2.4 Comparison of the processing time by using different algorithms

In our experiment, we compared the processing time of LBP, LGS, SLGS and MOW-SLGS, respectively, and they are shown in Table 4.

Table 4.The processing time of different algorithms

From Table 4, we can see that the processing time of LBP, LGS, SLGS and MOW-SLGS is 0.4764 s, 0.4764 s, 0.4969 s, 1.5517 s, respectively. The processing time for MOW-SLGS is highest, because it uses four directions. MOW-SLGS’s maximum processing time is almost four times that of SLGS.

4.2.5 Comparison of the performance of using the different algorithms

(I)Comparison between the weight and non-weight SLGS

In order to test the algorithm’s performance by modifying the weights and increasing multi-direction, we compared the recognition rate of SLGS, W-SLGS and MOW-SLGS with the same database.

The experimental results are shown in Table 5. When the number of training samples is 1, 4, or 5,W-SLGS improves somewhat on SLGS; however, when the number of samples is 2 or 3, W-SLGS does not do as well as SLGS.

Table 5.Recognition rate comparison between the weight and non-weight SLGS

When the number of samples is 2, 3, or 5, MOW-SLGS outperforms W-SLGS, but when the number of samples is 1 or 4, it does not even do as well as SLGS. However, overall, modifying weights and increasing multi-direction both improve the performance of SLGS algorithm.

(II) Comparison of the performance of different algorithms

This experiment compared the recognition rate of the LGS, SLGS, LBP, and MOW-SLGS methods under the same finger vein database, and for each set of experiments, we use the average of three repeated experiments as a result. The results are shown in Table 6. From the table, we can see that when the number of training samples (N) is 1, the recognition rate of the LGS, SLGS, LBP, and MOW-SLGS methods are 67.10%, 79.30%, 69.60%, and 81.30%, respectively. Of the algorithms, MOW-SLGS has the highest recognition rate. When N is 3, the recognition rate of the LGS, SLGS, LBP, and MOW-SLGS methods are 79.17%, 85.17%, 78.33%, and 87.17%, respectively, making MOW-SLGS again highest. When N is 5, the recognition rate of the LGS, SLGS, LBP, and MOW-SLGS methods are 94.50%, 95.00%, 95.00%, 96.00%, respectively, making all rates high, but MOW-SLGS still highest. MOW-SLGS thus has the best performance overall.

Table 6.Recognition rate of different algorithms under the same finger vein database

Since the finger vein image is rich in direction information, and LGS and SLGS only use the pixel information in the 0° direction of the target pixel, not that in the 45°, 90°, and 135° directions, they are less effective. The LBP operator uses only eight pixels around the target pixel and the magnitude relationship between the target pixel and the surrounding pixels, but ignores the gradient information between the surrounding pixels, costing it some feature extraction.

Fig. 14 shows the recognition rates of LGS, SLGS, PCA, LBP and MOW-SLGS more intuitively. As can be seen from it, the curve of MOW-SLGS is above other curves, and highest when the number of training samples is 5.

Fig. 14.Recognition rate of different algorithms under the same finger vein database

From another view, we draw the ROC curves [4] for LBP, LGS, SLGS, and MOW-SLGS, shown in Fig. 15 under the same finger vein database with the 1-NN classifier. Again, the curve of MOW-SLGS is on bottom of the other curves, which means that MOW-SLGS shows the best performance. The false acceptance rate (FAR) and false rejection rate (FRR) are evaluated with 2×(100× (100-1)×4) =79200 imposter matches versus 2×(100×4)=800 genuine matches respectively. In the experiment, we use Mahalanobis distance as the matching method.

Fig. 15.ROC curves of the different algorithms

Table 7 compares EER rate of different algorithms under the same finger vein database with the 1-NN classifier. As we can see from the table, the EER rate of LBP, LGS, SLGS, MOW-SLGS are 8.52%, 7.82%, 5.00%, 4.50% respectively, which means MOW-SLGS show better results compared with other algorithms.

Table 7.EER rate of different algorithms under the same finger vein database

 

5. Conclusion

This paper proposes a novel feature extraction (MOW-SLGS) method, and applies it to finger vein recognition. The MOW-SLGS method uses gradients and location of the pixels around target pixel, assigns weight to each edge according to the position relationship between the edge and the target pixel, and improves on SLGS in its resistance to changes in light and background factors. We test it with different classifiers (1-NN, ELM, VBELM). Experimental results show that the proposed method recognizes finger veins more easily than any other methods.